Games do pretty simple things on the GPU. Download textures, then transform them. They do a _lot_ of it, but conceptually the operations are not actually all that complicated or varied, by and large.

Browsers have this little thing they have to deal with called "text". Turns out that stuffing reasonable text into a texture and then scrolling it, say, is not something GPUs can do a good job of. This is why games just bitmap all their text and it looks like crap, by the way. I doubt you'd want text in a browser to look like it does in most games.

Browsers are working on making this faster, by the way. But it also turns out that not only are GPUs designed around stuff that games do (which is not the same set of stuff that browsers need done), but so are graphics drivers. The parts of them that Firefox 4 has been trying to use are clearly undertested, because they're ... rather buggy. Which is why there's now a driver blacklist (whitelist on Linux) and so forth...

One other thing. The expensive thing with shadows is not just the compositing, but recomputing the shadow itself. Any time something changes about the thing that's casting the shadow you have to rebuild the shadow. If the change is of a kind the GPU knows about (polygon moved or rotated or whatnot), it might be doable on the GPU. If it's something like "the width changed and the text layout has to be redone", you lose.

Because these calculations are done on software, not GPU. GPU programming is hard. Using the GPU requires extra-integration and extra-work that most developers are not willing to do.